Sugarcane parent database construction system based on clustering algorithm

Abstract

The invention relates to the technical field of agricultural biotechnology and intelligent breeding, in particular to a sugarcane parent database construction system based on a clustering algorithm, which comprises a data acquisition and preprocessing module, a clustering analysis module, a data cleaning module, a training set construction module, a model training module and a model integration and application module, genotype and phenotype data are processed in a standardized mode to form an analysis basis; analyzing a parent genetic structure by utilizing clustering and generating an initial class group; abnormal points are identified based on the class group center distance, missing phenotypes are filled, and data cleaning is completed; screening complete data parents in the center area of each type of group to construct a core training set; training a phenotype performance prediction model; finally, the model is integrated into a database, automatic prediction of the performance of the newly-added parents and the combined filial generations is achieved, and accurate data support is provided for breeding decision making.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of agricultural biotechnology and intelligent breeding technology, and particularly relates to a sugarcane parent database construction system based on a clustering algorithm. BACKGROUND

[0002] As one of the most important sugar and energy crops in the world, the improvement of sugarcane breeding efficiency is of great importance to national food security and sustainable agricultural development. Modern sugarcane breeding highly depends on the in-depth analysis and utilization of genotype and phenotype data of parent germplasm resources, and the construction of a structured parent database has become the basis for realizing data-driven breeding decisions. However, the sugarcane genome is highly complex, being polyploid and having a mixed genetic background, and the phenotype data accumulated in the breeding process is easily disturbed by the environment and has a large amount of missing and erroneous data, which makes it difficult for traditional databases to support the construction of high-precision prediction models.

[0003] The prior art lacks a systematic process that can effectively integrate genetic background analysis and data quality control, resulting in uneven quality of underlying data in the database, and further leading to low accuracy and insufficient reliability of genomic prediction models trained based on such data, which seriously restricts the precision and efficiency of parent selection and offspring screening in sugarcane breeding. SUMMARY

[0004] To make up for the above shortcomings, the present application provides a sugarcane parent database construction system based on a clustering algorithm, aiming to improve the problem in the prior art that there is a lack of a systematic process that can effectively integrate genetic background analysis and data quality control, resulting in uneven quality of underlying data in the database, and further leading to low accuracy and insufficient reliability of genomic prediction models trained based on such data.

[0005] In a first aspect, the present application provides the following technical solution: a sugarcane parent database construction system based on a clustering algorithm, comprising the following modules:

[0006] A data acquisition and preprocessing module is used to acquire genotype data, phenotype data and pedigree data of sugarcane parents, perform quality control and missing value filling on the genotype data to form a genotype numerical matrix, and perform environment effect correction on the phenotype data to obtain phenotypic genetic values.

[0007] A clustering analysis module is used to group all sugarcane parents based on the genotype numerical matrix using a clustering algorithm, generate initial clusters, and determine the optimal number of clusters.

[0008] The data cleaning module is used to calculate the distance from each sugarcane parent to the center of its initial group based on the initial group, thereby identifying genetic anomalies and performing manual verification and correction. At the same time, for parents with missing phenotypic data, the data is filled in based on the phenotypic genetic values ​​of other parents in their initial group, thereby obtaining a cleaned and corrected group.

[0009] The training set construction module is used to select parents with complete data located in the central region of each group from the cleaned and corrected groups, and to build a core training germplasm set.

[0010] The model training module is used to train a phenotypic performance prediction model based on the core training germplasm set and its corresponding genotype numerical matrix and phenotypic genetic value.

[0011] The model integration and application module is used to integrate the trained phenotypic performance prediction model into the database for phenotypic performance prediction of newly added sugarcane germplasm.

[0012] Preferably, the process for forming the genotype numerical matrix includes:

[0013] Obtain raw genotyping data obtained through SNP microarrays or simplified genome sequencing technology;

[0014] The original genotyping data is subjected to quality control, and SNP sites with a detection rate lower than a preset first threshold in all parents and a deletion rate higher than a preset second threshold in a single parent are removed, as are parental materials with a genotype data deletion rate higher than a preset third threshold, to obtain quality-controlled genotype data.

[0015] Each SNP site in the quality-controlled genotype data is numerically encoded, with the major allele homozygous type encoded as the first value, the heterozygous type encoded as the second value, and the minor allele homozygous type encoded as the third value, forming an initial numerical matrix.

[0016] Based on the initial numerical matrix, the K-nearest neighbor algorithm is used to calculate the filler value for each missing genotype data according to the genotype profile of other SNP loci, thereby generating a complete genotype numerical matrix.

[0017] Preferably, the process for obtaining the phenotypic genetic value includes:

[0018] Collect raw phenotypic data of sugarcane parents measured in field trials at multiple locations and / or in multiple years;

[0019] The genetic effects of the sugarcane parent itself were set as fixed effects, and the environmental effects and the interaction effects between the sugarcane parent and the environment were set as random effects, thus constructing a linear mixed model for phenotypic data analysis.

[0020] Solving the linear mixed model by using restricted maximum likelihood method, the genetic effect value of each sugarcane parent is estimated;

[0021] The estimated genetic effect value is output as the phenotype genetic value eliminating environmental interference for subsequent analysis.

[0022] Preferably, the generation process of the initial cluster includes:

[0023] Based on the genotype numerical matrix, the genetic distance between all sugarcane parents is calculated, and a genetic distance matrix is constructed;

[0024] Using K-means clustering algorithm, the genetic distance matrix is taken as input, and multiple iterations of clustering are performed within the preset K value range;

[0025] Based on the results obtained under each K value, the silhouette coefficient is calculated, and the K value corresponding to the maximum silhouette coefficient is determined as the optimal clustering number;

[0026] Using the determined optimal clustering number as the final K value, the K-means clustering algorithm is re-executed, and all sugarcane parents are divided into a corresponding number of stable initial clusters.

[0027] Preferably, the identification process of the genetic abnormal point includes:

[0028] For each generated initial cluster, the mean value of all parents in the cluster at each SNP site in the genotype numerical matrix is calculated to form a cluster center coordinate representing the genetic background of the cluster;

[0029] For each sugarcane parent, the Euclidean distance between its coordinate in the genotype numerical matrix and the cluster center coordinate of its belonging cluster is calculated, which is recorded as the intra-cluster distance;

[0030] Based on the distribution of the intra-cluster distance of all members in each initial cluster, an abnormality determination threshold is set for the cluster;

[0031] The parent with an intra-cluster distance greater than the abnormality determination threshold set by its belonging cluster is identified as a genetic abnormal point;

[0032] All identified genetic abnormal points are manually compared and verified with their pedigree records. For those with record errors or sample confusion, corrections or deletions are made in the database.

[0033] Preferably, the filling process according to the phenotype genetic value of other parents in its belonging initial cluster includes:

[0034] In the initial cluster, identify the parent with missing phenotype data as the target to be filled in;

[0035] For each target to be filled in, select all parents with complete phenotype data from the initial cluster to which it belongs to form an effective reference set;

[0036] Calculate the arithmetic mean of the phenotypic genetic values of all parents in the effective reference set for the missing phenotype trait, and fill the arithmetic mean into the missing phenotype data corresponding to the target to be filled in.

[0037] Preferably, the construction process of the core training germplasm set includes:

[0038] For each cleaned and corrected cluster, calculate the intra-cluster distance of all parents in the cluster to the cluster center, and select the parents with intra-cluster distance less than a set proportion threshold as candidate center parents;

[0039] Remove individuals with missing genotype data or phenotypic genetic values from the selected candidate center parents, and retain parents with complete data;

[0040] Collect all candidate center parents with complete data in all clusters to form the core training germplasm set.

[0041] Preferably, the training process of the phenotype performance prediction model includes:

[0042] Extract the genotype data of all parents in the core training germplasm set from the genotype value matrix as the model input feature matrix, and obtain the corresponding phenotypic genetic values of these parents as the model prediction target variable;

[0043] Select a genomic selection model as the training basis of the phenotype performance prediction model;

[0044] Input the model input feature matrix and the model prediction target variable into the genomic selection model, use cross-validation method to train the model, and optimize each parameter in the model;

[0045] When the model performance reaches a stable optimum under cross-validation, save the final parameter configuration and generate the trained phenotype performance prediction model.

[0046] Preferably, the process of predicting the phenotype performance of newly added sugarcane germplasm includes:

[0047] When a new sugarcane germplasm needs to be predicted, collect its genotype data; quality control the genotype data and remove SNP sites with detection rate below a set threshold;

[0048] The quality-controlled genotype data is numerically encoded, and missing genotype data is filled in to form an input genotype vector that is consistent with the characteristics of the genotype numerical matrix.

[0049] The obtained input genotype vector is input into the trained phenotypic performance prediction model;

[0050] The phenotypic performance prediction model is calculated based on the input genotype vector and outputs a phenotypic performance prediction value for the new sugarcane germplasm.

[0051] The predicted phenotypic performance values ​​are associated with the genotype information of the new germplasm and stored in a database to guide the breeding evaluation and selection decisions of the new germplasm.

[0052] Secondly, the present invention provides the following technical solution: a method for constructing a sugarcane parent database based on a clustering algorithm, the method comprising the following steps:

[0053] Genotype data, phenotypic data, and pedigree data of sugarcane parents were collected. The genotype data were subjected to quality control and missing value filling to form a genotype numerical matrix. The phenotypic data were corrected for environmental effects to obtain phenotypic genetic values.

[0054] Based on the genotype numerical matrix, a clustering algorithm was used to group all sugarcane parents, generate initial clusters, and determine the optimal number of clusters.

[0055] Based on the initial taxonomic group, the distance from each sugarcane parent to the center of its initial taxonomic group is calculated to identify genetic anomalies and perform manual verification and correction. Meanwhile, for parents with missing phenotypic data, the data is filled in based on the phenotypic genetic values ​​of other parents in their initial taxonomic group, thereby obtaining a cleaned and corrected taxonomic group.

[0056] Among the cleaned and corrected groups, parents with complete data located in the central region of each group are selected and a core training germplasm set is constructed.

[0057] Based on the core training germplasm set, its corresponding genotype numerical matrix and phenotypic genetic value, a phenotypic performance prediction model is trained.

[0058] The trained phenotypic performance prediction model is integrated into the database for predicting the phenotypic performance of newly added sugarcane germplasms.

[0059] The present invention has the following beneficial effects:

[0060] 1. In this invention, by organically integrating cluster analysis, data cleaning, and model training, a self-optimizing closed-loop system is constructed, which solves the defect of low model prediction accuracy caused by poor data quality in traditional databases. The system uses clustering results to objectively identify genetic anomalies and intelligently fill in missing phenotypes, which significantly improves the accuracy and completeness of the underlying data. Then, the cleaned high-quality data is used to construct a core training set, providing an effective data foundation for training a high-precision and high-reliability phenotypic performance prediction model.

[0061] 2. In this invention, through the prediction model completed by integrated training, breeders can quickly and accurately predict the potential phenotypic performance of new germplasm based solely on its genotype data, which greatly shortens the breeding cycle and reduces the cost and workload of field trials.

[0062] 3. In this invention, the systematic data processing flow ensures that newly added germplasm data can be automatically cleaned, standardized and included in the analysis, so that the database and prediction model can be continuously iterated and optimized as data accumulates, forming a dynamically growing knowledge base. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the architecture of a sugarcane parent database construction system based on clustering algorithm proposed in this invention;

[0064] Figure 2 This is a flowchart illustrating a method for constructing a sugarcane parent database based on a clustering algorithm, as proposed in this invention. Detailed Implementation

[0065] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Example 1:

[0067] In a first embodiment of the present invention, the present invention provides a sugarcane parent database construction system based on a clustering algorithm, such as... Figure 1 As shown, it includes the following modules:

[0068] The data acquisition and preprocessing module is used to collect genotype data, phenotypic data, and pedigree data of sugarcane parents, perform quality control and missing value imputation on genotype data to form a genotype numerical matrix, and perform environmental effect correction on phenotypic data to obtain phenotypic genetic values.

[0069] Furthermore, the process for forming the genotype numerical matrix includes:

[0070] Obtain raw genotyping data obtained through SNP microarrays or simplified genome sequencing technology;

[0071] The original genotyping data was quality controlled by removing SNP sites with a detection rate lower than the preset first threshold in all parents and a deletion rate higher than the preset second threshold in a single parent, as well as parental materials with a genotyping data deletion rate higher than the preset third threshold, to obtain the quality-controlled genotyping data.

[0072] Each SNP site in the quality-controlled genotype data is numerically encoded, with the major allele homozygous type encoded as the first value, the heterozygous type encoded as the second value, and the minor allele homozygous type encoded as the third value, forming an initial numerical matrix.

[0073] Based on the initial numerical matrix, the K-nearest neighbor algorithm is used to calculate the filler value for each missing genotype data according to the genotype profile of other SNP loci, generating a complete genotype numerical matrix.

[0074] Furthermore, the process for obtaining phenotypic genetic values ​​includes:

[0075] Collect raw phenotypic data of sugarcane parents measured in field trials at multiple locations and / or in multiple years;

[0076] The genetic effects of the sugarcane parent itself were set as fixed effects, while the environmental effects and the interaction effects between the sugarcane parent and the environment were set as random effects. A linear mixed model was constructed for phenotypic data analysis.

[0077] The restricted maximum likelihood method was used to solve the linear mixed model and estimate the genetic effect value of each sugarcane parent.

[0078] The estimated genetic effect values ​​are output as phenotypic genetic values ​​that have eliminated environmental interference, and used for subsequent analysis.

[0079] Specifically, genotypic data is obtained through commercial SNP microarrays (such as custom microarrays on the Illumina Infinium platform) or simplified genome sequencing technology. Phenotypic data comes from multi-year, multi-location field trials and includes at least three key traits: sugarcane stalk yield, sucrose content, and fiber content. Pedigree data is obtained by compiling historical breeding records. The system reads raw genotyping data (such as VCF files or microarray output files). The quality control process adopts the following standards: for SNP loci, loci with a detection rate of less than 95% in all sugarcane parents (i.e., the preset first threshold) are removed; for individual parents, loci with missing genotypic data are removed. For individuals with a loss rate higher than 20% (i.e., the preset second threshold), this step yields a quality-controlled genotype data list. The quality-controlled genotype data is then converted into a numerical matrix, with the following encoding rules: the major allele homozygous type (e.g., allele A, which is more frequent in the population, and its homozygous form AA) is encoded as 0 (i.e., the first value), the heterozygous type (e.g., AT or AG) is encoded as 1 (i.e., the second value), and the minor allele homozygous type (e.g., aa) is encoded as 2 (i.e., the third value). This forms an initial numerical matrix that represents the parental behavior and lists the SNP loci, still containing some missing values ​​(represented by NA).

[0080] The K-nearest neighbor algorithm was used to fill the missing values ​​in the initial numerical matrix. Specifically, the KNNImputer function from the scikit-learn library was used, with the number of neighbors n_neighbors=5. Euclidean distance was used to measure the genotypic similarity between parents. Based on the genotypic profile of each parent with a missing value at all other SNP loci, the algorithm found the 5 most similar neighboring parents. Then, the mean or mode of the genotypic values ​​of these 5 neighbors at that SNP locus was used for filling. After filling, a complete genotypic numerical matrix without missing values ​​was generated for all subsequent analyses.

[0081] Collect field phenotypic data from at least two locations over three consecutive years to ensure that each parent has repeated observation records in multiple environments. Use the lme4 package in R language to construct a linear mixture model. Taking sucrose content as an example, the model form is as follows: sucrose content ~ parental genotype + (1|environment) + (1|parental genotype:environment);

[0082] In this model, parental genotype is set as a fixed effect to estimate the genetic effect of each parent. Environment and parental genotype: Environment (i.e., the interaction between genotype and environment) are set as random effects. The restricted maximum likelihood method embedded in the lme4 package is used to solve the model to estimate the variance components of the random effects and the fixed effect value. From the model solution, the effect value of each parental genotype is extracted as the fixed effect value. This effect value is the best linear unbiased estimate. It removes the interference of environment and the interaction between genotype and environment and represents the heritable sucrose potential of the parent. The system outputs and stores this value as the phenotypic genetic value of the parent.

[0083] Through the above steps, the multi-source heterogeneous genotype and phenotype data can be standardized, denoised, and normalized, thus constructing a high-quality and highly comparable standardized data base for all subsequent analyses, fundamentally ensuring the accuracy and reliability of the final prediction model.

[0084] The clustering analysis module is used to group all sugarcane parents based on the genotype numerical matrix using a clustering algorithm, generate initial clusters, and determine the optimal number of clusters.

[0085] Furthermore, the process for generating the initial class group includes:

[0086] Based on the genotype numerical matrix, the genetic distances between all sugarcane parents are calculated, and a genetic distance matrix is ​​constructed.

[0087] The K-means clustering algorithm is used, with the genetic distance matrix as input, and multiple iterations of clustering are performed within a preset range of K values;

[0088] Based on the clustering results obtained at each K value, the silhouette coefficient is calculated, and the K value corresponding to the maximum silhouette coefficient is determined as the optimal number of clusters.

[0089] Using the determined optimal number of clusters as the final K value, the K-means clustering algorithm is re-executed to divide all sugarcane parents into a corresponding number of stable initial clusters.

[0090] Specifically, the system reads the genotype numerical matrix generated by the data acquisition and preprocessing module, and uses the Euclidean distance formula to calculate the genetic distance between all sugarcane parents. For two parents i and j, their genetic distance is calculated as follows: The calculation formula is:

[0091] ;

[0092] Where m is the total number of SNP sites, and These are the coding values ​​(0, 1, 2) of parents i and j at the k-th SNP locus, respectively. All pairwise distances are combined to construct an n×n symmetric genetic distance matrix (n is the total number of parents).

[0093] The K-means++ clustering algorithm is used, with the genetic distance matrix mentioned above as input. A reasonable search range for K values ​​is set, for example, from 2 to 10. For each K value within this range, the KMeans function from the scikit-learn library is used for multiple iterations to obtain stable clustering results. For each clustering result obtained under each K value, its silhouette coefficient is calculated. The silhouette coefficient measures the similarity of an object to its own cluster relative to other clusters, and its value range is [-1, 1]. The larger the value, the better the clustering effect. The system automatically compares the silhouette coefficients corresponding to all K values, and the K value corresponding to the global maximum silhouette coefficient is determined as the final optimal number of clusters. );

[0094] Using the optimal number of clusters determined above As the final parameter, the K-means++ clustering algorithm is re-executed once, this time dividing all sugarcane parents into... An initial taxonomic group is established, and the system assigns a taxonomic group label to each parent (e.g., 1, 2, 3, ...). The system calculates the center coordinates of each group (i.e., the mean vector of the encoded values ​​of all parents in each SNP position within the group). These stable initial groups and their center coordinates will be output and used as input to the data cleaning module.

[0095] Through the above steps, the inherent genetic structure of the parent population can be revealed through objective mathematical methods, providing a scientific basis for subsequent targeted data cleaning and core germplasm screening, and transforming the group cognition at the level of breeding experience into quantifiable classification criteria.

[0096] The data cleaning module is used to calculate the distance from each sugarcane parent to the center of its initial taxonomy based on the initial taxonomy, thereby identifying genetic anomalies and performing manual verification and correction. At the same time, for parents with missing phenotypic data, the data is filled in based on the phenotypic genetic values ​​of other parents in their initial taxonomy, thus obtaining a cleaned and corrected taxonomy.

[0097] Furthermore, the process for identifying genetic abnormalities includes:

[0098] For each initial taxonomic group generated, the mean value of all parents in the genotype numerical matrix at each SNP locus is calculated to form the taxonomic center coordinates representing the genetic background of the taxonomic group.

[0099] For each sugarcane parent, calculate its coordinates in the genotype numerical matrix and the Euclidean distance between it and the coordinates of the group center of its class, denoted as the intraclass distance;

[0100] Based on the distribution of intra-class distances of all members within each initial class group, an anomaly detection threshold is set for that class group.

[0101] Parents whose intraclass distance is greater than the anomaly threshold set for their class are identified as genetic anomalies.

[0102] All identified genetic anomalies will be manually compared and verified with their pedigree records. Those that are confirmed to be due to record errors or sample confusion will be corrected or removed from the database.

[0103] Furthermore, the process of filling in the genetic information based on the phenotypic values ​​of other parents in its initial taxonomy includes:

[0104] In the initial group, parents with missing phenotypic data were identified as targets to be filled;

[0105] For each target to be filled, all parents with complete phenotypic data are selected from its initial group to form a valid reference set;

[0106] Calculate the arithmetic mean of the phenotypic genetic values ​​of all parents in the valid reference set for the missing phenotypic trait, and fill the missing phenotypic data corresponding to the target to be filled with this arithmetic mean.

[0107] Specifically, the system reads the initial clustering results generated by the clustering analysis module, and for each cluster... Calculate the arithmetic mean of the values ​​of all parents at each SNP locus in the genotype numerical matrix within this group, forming a mean vector. The mean vector represents the coordinates of the group center, which symbolizes the typical genetic background of this group. The formula for calculating the j-th component is:

[0108] ;

[0109] in, It is a group The number of parents in the species, This represents the coding value of the i-th parent at the j-th SNP locus;

[0110] For each sugarcane parent i in the genotype numerical matrix, calculate its genotype coordinate vector. and its class center coordinates The Euclidean distance between the two parents is denoted as the intra-class distance of that parent. The calculation formula is as follows:

[0111] ;

[0112] Where m is the total number of SNP loci, this distance quantifies the degree of deviation of an individual from the core genetic background of its group;

[0113] The system performs separate tests on each initial group. To process this, firstly, calculate the mean intra-class distance of all members within the group. and standard deviation Then, set the anomaly detection threshold. Set as:

[0114] ;

[0115] Finally, in this group, those that satisfy... The parent i is automatically identified and marked as a genetic abnormality point;

[0116] The system provides an interactive interface that lists all identified genetic anomalies and their detailed pedigree information. Breeding experts can manually verify the data on this interface by comparing the genotype data of the anomalies with their pedigree records (paternal and maternal parents). For those confirmed to be pedigree record errors or field sample confusion, experts can directly correct them in the database or mark them for removal.

[0117] By executing in parallel with the above process, the system scans the initial group and identifies parents that are completely missing data in specific phenotypic traits (such as sucrose content), listing them as targets to be filled. For each target to be filled The system originates from its initial class. In the process, all parents with complete data on this phenotypic trait are selected, and these parents constitute a valid reference set. The parents in this set are closest to the genetic background of the target to be filled, providing the most reasonable reference basis. The above effective reference set is calculated. Phenotypic genotype values ​​for the missing trait in all parents. arithmetic mean The calculation formula is as follows:

[0118] ;

[0119] The system then uses this calculated average value. Fill to the target area The corresponding missing data location;

[0120] Through the above steps, the consistency of genetic background can be quantified for automated data quality inspection and repair, and the subjective and fragmented data cleaning work can be transformed into an objective and systematic standardized process, which significantly improves the intrinsic quality and consistency of the breeding database.

[0121] The training set construction module is used to select parents with complete data located in the central region of each group from the cleaned and corrected groups, and to build a core training germplasm set.

[0122] Furthermore, the construction process of the core training germplasm set includes:

[0123] For each cleaned and corrected taxa, calculate the intra-tax distance from all parents in the taxa to its taxa center, and select parents whose intra-tax distance is less than a set proportion threshold as candidate center parents.

[0124] Among the selected candidate central parents, individuals with missing genotype data or phenotypic genetic values ​​are removed, and parents with complete data are retained.

[0125] The candidate central parents with complete data from all taxa were collected to jointly construct a core training germplasm set.

[0126] Specifically, the system reads the cleaned and corrected data groups output by the data cleaning module, and for each data group... Perform the following operations: retrieve the group center coordinates of the group that have been calculated in the data cleaning module. and the intra-class distance of each parent i Calculate this class group The third quartile Q3 and the first quartile Q1 of the intraclass distance of all parents are used to obtain the interquartile range IQR = Q3 − Q1. Then, the screening threshold is set to Q1 − 1.5 × IQR − 1.5 × IQR. This threshold is used to identify the most representative parents distributed in the middle.

[0127] Groups All of the above satisfy Parent i is initially selected as candidate central parents. These parents are genetically closest to the core of the taxonomic group and best represent the typical characteristics of the taxonomic group. The system performs data integrity checks on all candidate central parents selected in the above steps. The checks include whether there are missing values ​​in the corresponding rows of the genotype numerical matrix, and whether they have valid phenotypic genetic values ​​for all target traits. Individuals with missing data in any of these areas are removed from the candidate list, and only parents with complete data are retained. Then, the system traverses all the cleaned and corrected taxonomic groups and gathers all the candidate central parents with complete data that have passed the above verification in each taxonomic group. This set, consisting of the most representative parents with the highest data quality in each taxonomic group, is finally defined as the core training germplasm set. This set will be output and used as input to the model training module to train a high-precision phenotypic performance prediction model.

[0128] Through the above steps, core germplasm with typical genetic background and complete data can be selected from various groups, constructing a high-quality and highly representative training set, which provides the optimal data foundation for subsequent training of prediction models with high accuracy and strong generalization ability.

[0129] The model training module is used to train a phenotypic performance prediction model based on the core training germplasm set and its corresponding genotype numerical matrix and phenotypic genetic values.

[0130] Furthermore, the training process for the phenotypic performance prediction model includes:

[0131] Genotype data of all parents in the core training germplasm set are extracted from the genotype numerical matrix and used as the model input feature matrix. At the same time, the phenotypic genetic values ​​of these parents are obtained as the model prediction target variables.

[0132] A genomic selection model was chosen as the training basis for the phenotypic performance prediction model.

[0133] The model input feature matrix and the model prediction target variable are input into the genome selection model, and the model is trained using cross-validation and the various parameters within the model are optimized.

[0134] When the model performance reaches a stable optimal level under cross-validation, the final parameter configuration is saved, and the trained phenotypic performance prediction model is generated.

[0135] Specifically, the system reads the core training germplasm set list generated by the training set construction module. Based on this list, it extracts all rows corresponding to the parents from the global genotype numerical matrix to form a sub-matrix. Each row of this sub-matrix contains all SNP genotype data of a parent, and each column contains a SNP locus. This matrix is ​​the feature matrix X input to the model, with dimensions n×p, where n is the number of core training germplasms and p is the number of SNP loci. At the same time, it extracts the values ​​of these core training germplasms on specific target traits (such as sucrose content) from the stored phenotypic genetic value data to form a vector of length n. This vector is the target variable y predicted by the model.

[0136] In this embodiment, the GBLUP model in genome selection is preferred as the training basis for the phenotypic performance prediction model. This model assumes that all marker effects follow a normal distribution and captures the genetic similarity between individuals by constructing a genome relation matrix. The formula for calculating the genome relation matrix G is as follows:

[0137] ;

[0138] Where W is the centered genotype matrix (i.e., the mean of the column in which each element of the feature matrix X is subtracted). Allelic frequency of the j-th SNP;

[0139] The prepared feature matrix X and target variable y are input into the GBLUP model. The model is trained and evaluated using the K-fold cross-validation method, with K=5. The specific process is as follows: the dataset is randomly divided into 5 non-overlapping subsets. Each subset is used as the validation set, and the other 4 subsets are used as the training set. The genome relationship matrix is ​​constructed on the training set and the model is solved. Then, predictions are made on the validation set.

[0140] The model solution uses the restricted maximum likelihood method to estimate the variance components (i.e., genetic variance). and residual variance Then, the effect value of each marker is calculated or the genomic estimated breeding value of an individual is obtained directly. This process itself is an optimization process of the key parameters (variance components) within the model, aiming to find the parameter combination that best explains the genetic structure of the current data.

[0141] After cross-validation, the system calculates the average prediction performance of the model across all validation sets (usually using the prediction correlation coefficient or mean squared error as an indicator). When the model performance remains stable and reaches its optimal level in successive cross-validation rounds, the model is retrained once using all core training germplasm datasets, and the final model parameters are saved, including but not limited to the genomic relation matrix and estimated variance components. and And the final marker effect value or the estimated breeding value of the genome of all training individuals, at this point, a phenotypic performance prediction model that has been trained and can be used to predict the phenotypic performance of new germplasm is generated.

[0142] Through the above steps, a high-quality core training set can be used to encapsulate the complex relationship between genotype and phenotype into a computable predictive model. This enables the core function of accurately predicting the breeding value of untested parents based solely on genotype data, providing a quantitative tool for precise mating.

[0143] The model integration and application module is used to integrate the trained phenotypic performance prediction model into the database for phenotypic performance prediction of newly added sugarcane germplasm.

[0144] Furthermore, the process for predicting the phenotypic performance of newly added sugarcane germplasm includes:

[0145] When new sugarcane germplasm needs to be predicted, its genotype data is collected; the genotype data is quality controlled, and SNP sites with a detection rate lower than the set threshold are removed.

[0146] The quality-controlled genotype data is numerically encoded, and missing genotype data is filled in to form an input genotype vector that is consistent with the characteristics of the genotype numerical matrix.

[0147] The obtained input genotype vector is then input into the trained phenotypic performance prediction model.

[0148] The phenotypic performance prediction model is calculated based on the input genotype vector and outputs a phenotypic performance prediction value for the new sugarcane germplasm.

[0149] The predicted phenotypic performance values ​​are correlated with the genotype information of the new germplasm and stored in a database to guide the breeding evaluation and selection decisions of the new germplasm.

[0150] Specifically, when a user submits a new sugarcane germplasm resource (hereinafter referred to as new germplasm) and its genotype data (such as VCF file) through the system interface, the system automatically triggers the following preprocessing process. The system calls the same quality control standards and algorithms as described above. Specifically, it uses the SNP locus list corresponding to the trained model as a benchmark. If the detection rate of a certain SNP locus in the new germplasm data is lower than 95% (i.e., the set threshold), the locus is directly removed. This step ensures that the SNP locus set of the input model is completely consistent with the training set. Subsequently, the system calls the same encoding rules and filling model as described above. First, the quality-controlled genotype data is encoded into numerical values ​​according to the same rules (0 for homozygous major alleles, 1 for heterozygous alleles, and 2 for homozygous minor alleles) to form an initial numerical vector. Then, the system loads the KNN missing value filling model that has been trained during the database construction phase and uses this model to fill in the missing values ​​that may exist in the initial vector, finally generating a complete input genotype vector that is completely consistent with the genotype numerical matrix features during the training phase.

[0151] The system inputs the preprocessed genotype vector into a pre-trained phenotypic performance prediction model (such as the GBLUP model) integrated into the database. The model internally calculates the genotype vector of the new germplasm based on its stored final parameters (such as marker effect values ​​or genomic relationship structures). For the GBLUP model, the predicted value is... The calculation is based on the following core idea: ,in It is a new germplasm genotype vector that has undergone centralized processing. It is the vector of labeled effect values ​​estimated during the model training phase. After the model is calculated, it outputs a specific, quantified predicted value of phenotypic performance.

[0152] The system automatically associates the predicted phenotypic performance value with the unique identifier and genotype data of the new germplasm, and stores it as a new record in a specific data table in the database. Users can query the prediction results through the system interface. The results directly serve breeding decisions, such as guiding hybridization (using new germplasm with excellent predicted performance as parental candidates), optimizing resource allocation (prioritizing field trials of materials with high predicted values), and early selection (eliminating individuals with poor predicted performance during the seedling stage), thereby achieving precise guidance for the breeding process.

[0153] Through the above steps, the offline-trained prediction model can be transformed into an online decision-making tool, enabling real-time, batch, and automated evaluation of the phenotypic performance of new germplasm, and improving the efficiency and foresight of parent evaluation and screening in the breeding process.

[0154] Example 2:

[0155] Existing methods lack a systematic process to effectively integrate genetic background analysis and data quality control, resulting in inconsistent quality of the underlying database data. This further leads to low accuracy and insufficient reliability of genome prediction models trained on such data. This deficiency severely restricts the accuracy and efficiency of parent selection and progeny screening in sugarcane breeding. To address these issues, this invention provides a method for constructing a sugarcane parent database based on a clustering algorithm, the structure of which is as follows: Figure 2 As shown. The specific implementation process of this method is as follows:

[0156] Genotypic, phenotypic, and pedigree data of sugarcane parents were collected. The genotypic data were subjected to quality control and missing value filling to form a genotypic numerical matrix. The phenotypic data were corrected for environmental effects to obtain phenotypic genetic values.

[0157] Based on the genotype numerical matrix, a clustering algorithm was used to group all sugarcane parents, generate initial clusters, and determine the optimal number of clusters.

[0158] Based on the initial taxonomic group, the distance from each sugarcane parent to the center of its initial taxonomic group is calculated to identify genetic abnormalities and perform manual verification and correction. At the same time, for parents with missing phenotypic data, the data is filled in based on the phenotypic genetic values ​​of other parents in their initial taxonomic group, thereby obtaining a cleaned and corrected taxonomic group.

[0159] After cleaning and correction, parents with complete data located in the central region of each group are selected and a core training germplasm set is constructed.

[0160] Based on the core training germplasm set, its corresponding genotype numerical matrix and phenotypic genetic values, a phenotypic performance prediction model is trained.

[0161] The trained phenotypic performance prediction model is integrated into the database for predicting the phenotypic performance of newly added sugarcane germplasms.

[0162] Specifically, firstly, the system collects multi-source data and performs standardized preprocessing to form a high-quality genotype numerical matrix and environmentally corrected phenotypic genetic values. Then, based on the genetic background, it performs cluster analysis on the parents to generate initial clusters. Subsequently, the cluster structure guides data cleaning, accurately identifies and corrects genetic anomalies, and intelligently fills in missing phenotypes based on genetic similarity within the clusters, obtaining high-quality cleaned and corrected clusters. On this basis, core germplasm with both genetic representativeness and data integrity is selected from each cluster to construct a high-quality training set. Based on this training set, a high-precision phenotypic performance prediction model is trained. Finally, the model is integrated into the database to achieve real-time prediction of the phenotypic performance of newly added germplasm, thus completing a closed loop from data aggregation, quality improvement, knowledge mining to intelligent decision-making.

[0163] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A sugarcane parent database construction system based on clustering algorithm, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to acquire genotype data, phenotypic data and pedigree data of sugarcane parents, perform quality control and missing value filling on the genotype data to form a genotype numerical matrix, and perform environmental effect correction on the phenotypic data to obtain phenotypic genetic values. The clustering analysis module is used to group all sugarcane parents based on the genotype numerical matrix using a clustering algorithm, generate initial clusters, and determine the optimal number of clusters. The data cleaning module is used to calculate the distance from each sugarcane parent to the center of its initial group based on the initial group, thereby identifying genetic anomalies and performing manual verification and correction. At the same time, for parents with missing phenotypic data, the data is filled in based on the phenotypic genetic values ​​of other parents in their initial group, thereby obtaining a cleaned and corrected group. The training set construction module is used to select parents with complete data located in the central region of each group from the cleaned and corrected groups, and to build a core training germplasm set. The model training module is used to train a phenotypic performance prediction model based on the core training germplasm set and its corresponding genotype numerical matrix and phenotypic genetic value. The model integration and application module is used to integrate the trained phenotypic performance prediction model into the database for phenotypic performance prediction of newly added sugarcane germplasm.

2. The sugarcane parent database construction system based on clustering algorithm according to claim 1, characterized in that, The process for forming the genotype numerical matrix includes: Obtain raw genotyping data obtained through SNP microarrays or simplified genome sequencing technology; The original genotyping data is subjected to quality control, and SNP sites with a detection rate lower than a preset first threshold in all parents and a deletion rate higher than a preset second threshold in a single parent are removed, as are parental materials with a genotype data deletion rate higher than a preset third threshold, to obtain quality-controlled genotype data. Each SNP site in the quality-controlled genotype data is numerically encoded, with the major allele homozygous type encoded as the first value, the heterozygous type encoded as the second value, and the minor allele homozygous type encoded as the third value, forming an initial numerical matrix. Based on the initial numerical matrix, the K-nearest neighbor algorithm is used to calculate the filler value for each missing genotype data according to the genotype profile of other SNP loci, thereby generating a complete genotype numerical matrix.

3. The sugarcane parent database construction system based on clustering algorithm according to claim 1, characterized in that, The process for obtaining the phenotypic genetic values ​​includes: Collect raw phenotypic data of sugarcane parents measured in field trials at multiple locations and / or in multiple years; The genetic effects of the sugarcane parent itself were set as fixed effects, and the environmental effects and the interaction effects between the sugarcane parent and the environment were set as random effects, thus constructing a linear mixed model for phenotypic data analysis. The linear mixed model was solved using the restricted maximum likelihood method to estimate the genetic effect value of each sugarcane parent. The estimated genetic effect value is output as the phenotypic genetic value after eliminating environmental interference, for subsequent analysis.

4. The sugarcane parent database construction system based on clustering algorithm according to claim 1, characterized in that, The process for generating the initial class group includes: Based on the genotype numerical matrix, calculate the genetic distance between all sugarcane parents and construct a genetic distance matrix; The K-means clustering algorithm is used, with the genetic distance matrix as input, and multiple iterations of clustering are performed within a preset range of K values; Based on the clustering results obtained at each K value, the silhouette coefficient is calculated, and the K value corresponding to the maximum silhouette coefficient is determined as the optimal number of clusters. Using the determined optimal number of clusters as the final K value, the K-means clustering algorithm is re-executed to divide all sugarcane parents into a corresponding number of stable initial clusters.

5. A sugarcane parent database construction system based on clustering algorithm according to claim 1, characterized in that, The process for identifying genetic abnormalities includes: For each initial taxonomic group generated, the mean value of all parents in the genotype numerical matrix at each SNP locus is calculated to form the taxonomic center coordinates representing the genetic background of the taxonomic group. For each sugarcane parent, calculate its coordinates in the genotype numerical matrix and the Euclidean distance between it and the coordinates of the group center of its class, denoted as the intraclass distance; Based on the distribution of intra-class distances of all members within each initial class group, an anomaly detection threshold is set for that class group. Parents whose intraclass distance is greater than the anomaly threshold set for their class are identified as genetic anomalies. All identified genetic anomalies will be manually compared and verified with their pedigree records. Those that are confirmed to be due to record errors or sample confusion will be corrected or removed from the database.

6. The sugarcane parent database construction system based on clustering algorithm according to claim 1, characterized in that, The process of filling in the genetic information based on the phenotypic values ​​of other parents in its initial taxonomy includes: In the initial group, parents with missing phenotypic data were identified as targets to be filled; For each target to be filled, all parents with complete phenotypic data are selected from its initial group to form a valid reference set; Calculate the arithmetic mean of the phenotypic genetic values ​​of all parents in the effective reference set for the missing phenotypic trait, and fill the missing phenotypic data corresponding to the target to be filled with this arithmetic mean.

7. The sugarcane parent database construction system based on clustering algorithm according to claim 1, characterized in that, The construction process of the core training germplasm set includes: For each cleaned and corrected taxa, calculate the intra-tax distance from all parents in the taxa to its taxa center, and select parents whose intra-tax distance is less than a set proportion threshold as candidate center parents. Among the selected candidate central parents, individuals with missing genotype data or phenotypic genetic values ​​are removed, and parents with complete data are retained. The candidate central parents with completely complete data from all taxa are collected to jointly construct the core training germplasm set.

8. A sugarcane parent database construction system based on clustering algorithm according to claim 1, characterized in that, The training process for the phenotypic performance prediction model includes: Genotype data of all parents in the core training germplasm set are extracted from the genotype numerical matrix and used as the model input feature matrix. At the same time, the phenotypic genetic values ​​of these parents are obtained as the model prediction target variables. A genomic selection model was selected as the training basis for the phenotypic performance prediction model. The model input feature matrix and the model prediction target variable are input into the genome selection model, and the model is trained using cross-validation and the various parameters within the model are optimized. When the model performance reaches a stable optimal level under cross-validation, the final parameter configuration is saved, and the trained phenotypic performance prediction model is generated.

9. A sugarcane parent database construction system based on clustering algorithm according to claim 1, characterized in that, The process for predicting the phenotypic performance of newly added sugarcane germplasm includes: When new sugarcane germplasm needs to be predicted, its genotype data is collected; the genotype data is quality controlled, and SNP sites with a detection rate lower than the set threshold are removed. The quality-controlled genotype data is numerically encoded, and missing genotype data is filled in to form an input genotype vector that is consistent with the characteristics of the genotype numerical matrix. The obtained input genotype vector is input into the trained phenotypic performance prediction model; The phenotypic performance prediction model is calculated based on the input genotype vector and outputs a phenotypic performance prediction value for the new sugarcane germplasm. The predicted phenotypic performance values ​​are associated with the genotype information of the new germplasm and stored in a database to guide the breeding evaluation and selection decisions of the new germplasm.

10. A method for constructing a sugarcane parent database based on a clustering algorithm, characterized in that, The method for constructing a sugarcane parent database based on a clustering algorithm, as described in any one of claims 1-9, comprises the following steps: Genotype data, phenotypic data, and pedigree data of sugarcane parents were collected. The genotype data were subjected to quality control and missing value filling to form a genotype numerical matrix. The phenotypic data were corrected for environmental effects to obtain phenotypic genetic values. Based on the genotype numerical matrix, a clustering algorithm was used to group all sugarcane parents, generate initial clusters, and determine the optimal number of clusters. Based on the initial taxonomic group, the distance from each sugarcane parent to the center of its initial taxonomic group is calculated to identify genetic anomalies and perform manual verification and correction. Meanwhile, for parents with missing phenotypic data, the data is filled in based on the phenotypic genetic values ​​of other parents in their initial taxonomic group, thereby obtaining a cleaned and corrected taxonomic group. Among the cleaned and corrected groups, parents with complete data located in the central region of each group are selected and a core training germplasm set is constructed. Based on the core training germplasm set, its corresponding genotype numerical matrix and phenotypic genetic value, a phenotypic performance prediction model is trained. The trained phenotypic performance prediction model is integrated into the database for predicting the phenotypic performance of newly added sugarcane germplasms.

Images